🎙 Song Forge

Type a vibe. Get a whole song.

No big-tech AI cloud, no subscription, no API keys. Songs render on my own hardware.

Get Song Forge on iPhone or Android: free, 10 songs on the house. No subscription, no account, no email.

Song Forge turns a one-line description into a finished song with lyrics, music and vocals.

Song Forge is a native iPhone (Apple) and Android (Google) app. Describe a song, like "sunset reggae with steel pan, 78 bpm", and Song Forge writes the lyrics, composes the music and vocals, and hands you a finished track. Drop in a 10-second voice sample and it'll re-sing the song in that voice. The phone sends your prompt to my own Mac render server (see PRIVACY.md); no third-party AI API is involved.

• Render server ios/forge_server.py: a standard-library Python HTTP server that takes an idea, writes lyrics, queues the song on ACE-Step, saves and tags the result, and serves the web UI and JSON API (/api/song, /api/status, /api/songs, /api/voices, ...).
• Lyric writing with guardrails (_llm_lyrics in forge_server.py): prompts a local Gemma model, rejects output containing banned phrases, retries, and falls back to generated seed lyrics.
• Karaoke lyric timing (_align_lyrics): runs Whisper for word timestamps, then fuzzy-matches the written lyrics to what was actually sung, so the on-screen lines stay in sync without showing Whisper's mishearings.
• Voice-swap pipeline (_run_swap_impl): Demucs vocal split, seed-vc voice conversion, then an ffmpeg remix, run as a background job with stage-aware progress.
• Genre-aware vocal prompting: per-genre vocal cues and auto voice assist for gospel, soul, hip-hop, reggae and more, to push back on the model defaulting every genre to the same pop voice.
• Web UI ios/index.html and the audio-reactive Three.js visualizer ios/viz3d.js.
• Launch tooling ios/forge_supervisor.sh and ios/launch.applescript: boots ACE-Step and the server idempotently and opens the UI.
• Android app MainActivity.kt: WebView shell with a remote URL pointer file and fallback domain, per-install random ID, a JS bridge shim matching the iOS webkit.messageHandlers convention, and native song downloads through DownloadManager. R8 keep rules for the JS bridge are in proguard-rules.pro.

Upstream, not mine: ACE-Step (music), seed-vc (voice conversion), Whisper (timing), Gemma (lyrics, served through LM Studio), MLX, Demucs, ffmpeg and Three.js. See CREDITS.md.

This repo holds both platforms, cleanly separated:

📱 iOS

The local Mac render engine (forge_server.py, forge_supervisor.sh) that powers the apps, its web UI, and App Store screenshots in ios/docs/appstore/. Source in ios/. The native iPhone wrapper (Swift) is not in this repo.

Status: ✅ Live on the App Store and Google Play: iPhone · Android.

Running the render server yourself (what the code expects):

• An Apple Silicon Mac. Paths are hardcoded to ~/Desktop/PROJECTS/Song Forge in the supervisor and AppleScript.
• ACE-Step 1.5 checked out at engines/ACE-Step-1.5 (API on port 8001) and seed-vc at engines/seed-vc with its own .venv. Engines are not included (see .gitignore).
• LM Studio serving Gemma at http://127.0.0.1:1234, plus ffmpeg, demucs and mlx_whisper installed.
• Then run bash ios/forge_supervisor.sh (or python3 ios/forge_server.py directly) and open http://127.0.0.1:8767/.

If LM Studio is down, lyrics fall back to simple generated seed lyrics. Publishing to the public site relies on a private ~/Scripts/songs_sync.py that is not in this repo, so auto-publish is disabled without it.

🤖 Android

The Android port. Source in android/. Build with cd android && ./gradlew assembleDebug (JDK 17, compileSdk 35).

Status: ✅ Live on Google Play.

Not done yet: in-app purchases on Android. The buy() bridge in MainActivity.kt is a stub that shows a "coming soon" notice; Google Play Billing is not wired up.

Links

• Product page: https://nicedreamzwholesale.com/software/song-forge/
• Questions or bugs? info@nicedreamzwholesale.com